Concavex Test Run
Paul Manser
September 20, 2016
Head Ache Data Example
Read data & initialize Gibbs sampler
library(rjags)
library(coda)
# read in data ------------------------------------------------------------
dose.orig <- log( 2 * c(0, 2.5, 5, 10, 20, 50, 100, 200) + 1)
dose <- dose.orig / max(dose.orig)
pred.doses <- seq(0, 1, length.out = 500)
mu.hat <- c(-2.22, -1.95, -2.05, -1.08, -1.45, -1.29, -1.17, -.57)
var.hat <- c(0.085, 0.286, 0.226, 0.084, 0.103, 0.091, 0.094, 0.075)
# initialize 3 Gibbs Sampling chains --------------------------------------
inits1 <- inits2 <- inits3 <- list(theta_0 = 0, theta_1 = 0, lambda = 0)
inits <- list(inits1, inits2, inits3)
# initialize model --------------------------------------------------------
setwd('bug')
jags.model <- jags.model('concavex-model.bug',
data = list('dose' = dose,
'eff' = mu.hat,
'tau' = 1 / var.hat, # jags deals in precisions
'n.arms' = length(dose),
'pred.doses' = pred.doses,
'pred.length' = length(pred.doses)),
inits = inits, n.chains = 3)## Compiling model graph
## Resolving undeclared variables
## Allocating nodes
## Graph information:
## Observed stochastic nodes: 8
## Unobserved stochastic nodes: 3
## Total graph size: 9624
##
## Initializing modelsetwd('..')Look at posteriors
jags.samples <- jags.samples(jags.model, c("theta_0", "theta_1", "lambda", "mu.tilde"), n.iter = 5000)
par(mfrow = c(1, 3))
hist(jags.samples$lambda, main = "posterior for lambda")
hist(jags.samples$theta_0, main = "posterior for theta_0")
hist(jags.samples$theta_1, main = "posterior for theta_1")
BGR Diagnostics
coda.samples <- coda.samples(jags.model, c("theta_0", "theta_1", "lambda"), n.iter = 5000)
gelman.plot(coda.samples)
Gibbs sampler trace
Use few iterations to not overwhelm vector graphics in pdf
coda.samples <- coda.samples(jags.model, c("theta_0", "theta_1", "lambda"), n.iter = 500)
par(mfrow=c(1, 3))
traceplot(coda.samples, smooth = TRUE)
Plot concavex fit w/ credible interval
dose.range <- seq(0, max(dose), length.out = 500)
dose.range.orig <- seq(0, max(dose.orig), length.out = 500)
theta_0_mean <- mean(jags.samples$theta_0)
theta_1_mean <- mean(jags.samples$theta_1)
lambda_mean <- mean(jags.samples$lambda)
# plot point estimates and curve using posterior means
par(mfrow = c(1,1))
plot(dose.orig, mu.hat, pch = 16, ylim = range(mu.hat) + c(-1, 1))
y <- theta_0_mean + theta_1_mean * (((((1-lambda_mean)/2)^2 + lambda_mean)/(((1-lambda_mean)/2)^2 + lambda_mean*dose.range))*dose.range)
lines(dose.range.orig, y, lwd = 2)
# add upper and lower 90% credible interval for curve
pci.ul <- pci.ll <- numeric(length(pred.doses))
for(ii in 1:length(pred.doses)) {
# quantiles <- quantile(unlist(jags.samples$y.pred[ii, , ]), probs = c(0.025, .975))
quantiles <- quantile(unlist(jags.samples$mu.tilde[ii, , ]), probs = c(0.05, .95))
pci.ll[ii] <- quantiles[1]
pci.ul[ii] <- quantiles[2]
}
lines(dose.range.orig, pci.ll, col = 'grey', lwd = 2)
lines(dose.range.orig, pci.ul, col = 'grey', lwd = 2)
rm(list=ls())Regeneron Example
Read data & initialize Gibbs sampler
# read in data ------------------------------------------------------------
dose.orig <- c(0, 1, 3, 6, 9) / 9
dose <- dose.orig / max(dose.orig)
pred.doses <- seq(0, 1, length.out = 500)
mu.hat <- c(-2.25, -3.35, -3.33, -3.03, -3.65)
var.hat <- rep(2.5^2, 5) / c(83, 85, 84, 85, 84)
# initialize 3 Gibbs Sampling chains --------------------------------------
inits1 <- inits2 <- inits3 <- list(theta_0 = 0, theta_1 = 0, lambda = 0)
inits <- list(inits1, inits2, inits3)
# initialize model --------------------------------------------------------
setwd('bug')
jags.model <- jags.model('concavex-model.bug',
data = list('dose' = dose,
'eff' = mu.hat,
'tau' = 1 / var.hat, # jags deals in precisions
'n.arms' = length(dose),
'pred.doses' = pred.doses,
'pred.length' = length(pred.doses)),
inits = inits, n.chains = 3)## Compiling model graph
## Resolving undeclared variables
## Allocating nodes
## Graph information:
## Observed stochastic nodes: 5
## Unobserved stochastic nodes: 3
## Total graph size: 9579
##
## Initializing modelsetwd('..')Look at posteriors
jags.samples <- jags.samples(jags.model, c("theta_0", "theta_1", "lambda", "mu.tilde"), n.iter = 5000)
par(mfrow = c(1, 3))
hist(jags.samples$lambda, main = "posterior for lambda")
hist(jags.samples$theta_0, main = "posterior for theta_0")
hist(jags.samples$theta_1, main = "posterior for theta_1")
BGR Diagnostics
coda.samples <- coda.samples(jags.model, c("theta_0", "theta_1", "lambda"), n.iter = 5000)
gelman.plot(coda.samples)
Gibbs sampler trace
Use few iterations to not overwhelm vector graphics in pdf
coda.samples <- coda.samples(jags.model, c("theta_0", "theta_1", "lambda"), n.iter = 500)
par(mfrow=c(1, 3))
traceplot(coda.samples, smooth = TRUE)
Plot concavex fit w/ credible interval
dose.range <- seq(0, max(dose), length.out = 500)
dose.range.orig <- seq(0, max(dose.orig), length.out = 500)
theta_0_mean <- mean(jags.samples$theta_0)
theta_1_mean <- mean(jags.samples$theta_1)
lambda_mean <- mean(jags.samples$lambda)
# plot point estimates and curve using posterior means
par(mfrow = c(1,1))
plot(dose.orig, mu.hat, pch = 16, ylim = range(mu.hat) + c(-1, 1))
y <- theta_0_mean + theta_1_mean * (((((1-lambda_mean)/2)^2 + lambda_mean)/(((1-lambda_mean)/2)^2 + lambda_mean*dose.range))*dose.range)
lines(dose.range.orig, y, lwd = 2)
# add upper and lower 90% credible interval for curve
pci.ul <- pci.ll <- numeric(length(pred.doses))
for(ii in 1:length(pred.doses)) {
# quantiles <- quantile(unlist(jags.samples$y.pred[ii, , ]), probs = c(0.025, .975))
quantiles <- quantile(unlist(jags.samples$mu.tilde[ii, , ]), probs = c(0.05, .95))
pci.ll[ii] <- quantiles[1]
pci.ul[ii] <- quantiles[2]
}
lines(dose.range.orig, pci.ll, col = 'grey', lwd = 2)
lines(dose.range.orig, pci.ul, col = 'grey', lwd = 2)
Null Regeneron Example
Read data & initialize Gibbs sampler
# read in data ------------------------------------------------------------
dose.orig <- c(0, 1, 3, 6, 9) / 9
dose <- dose.orig / max(dose.orig)
pred.doses <- seq(0, 1, length.out = 500)
mu.hat <- c(-2.25, -3.35, -3.33, -3.03, -3.65)
mu.hat[(-1)] <- mu.hat[(-1)] - mean(mu.hat[(-1)]) + mu.hat[1]
var.hat <- rep(2.5^2, 5) / c(83, 85, 84, 85, 84)
# initialize 3 Gibbs Sampling chains --------------------------------------
inits1 <- inits2 <- inits3 <- list(theta_0 = 0, theta_1 = 0, lambda = 0)
inits <- list(inits1, inits2, inits3)
# initialize model --------------------------------------------------------
setwd('bug')
jags.model <- jags.model('concavex-model.bug',
data = list('dose' = dose,
'eff' = mu.hat,
'tau' = 1 / var.hat, # jags deals in precisions
'n.arms' = length(dose),
'pred.doses' = pred.doses,
'pred.length' = length(pred.doses)),
inits = inits, n.chains = 3)## Compiling model graph
## Resolving undeclared variables
## Allocating nodes
## Graph information:
## Observed stochastic nodes: 5
## Unobserved stochastic nodes: 3
## Total graph size: 9579
##
## Initializing modelsetwd('..')Look at posteriors
jags.samples <- jags.samples(jags.model, c("theta_0", "theta_1", "lambda", "mu.tilde"), n.iter = 5000)
par(mfrow = c(1, 3))
hist(jags.samples$lambda, main = "posterior for lambda")
hist(jags.samples$theta_0, main = "posterior for theta_0")
hist(jags.samples$theta_1, main = "posterior for theta_1")
BGR Diagnostics
coda.samples <- coda.samples(jags.model, c("theta_0", "theta_1", "lambda"), n.iter = 5000)
gelman.plot(coda.samples)
Gibbs sampler trace
Use few iterations to not overwhelm vector graphics in pdf
coda.samples <- coda.samples(jags.model, c("theta_0", "theta_1", "lambda"), n.iter = 500)
par(mfrow=c(1, 3))
traceplot(coda.samples, smooth = TRUE)
Plot concavex fit w/ credible interval
dose.range <- seq(0, max(dose), length.out = 500)
dose.range.orig <- seq(0, max(dose.orig), length.out = 500)
theta_0_mean <- mean(jags.samples$theta_0)
theta_1_mean <- mean(jags.samples$theta_1)
lambda_mean <- mean(jags.samples$lambda)
# plot point estimates and curve using posterior means
par(mfrow = c(1,1))
plot(dose.orig, mu.hat, pch = 16, ylim = range(mu.hat) + c(-1, 1))
y <- theta_0_mean + theta_1_mean * (((((1-lambda_mean)/2)^2 + lambda_mean)/(((1-lambda_mean)/2)^2 + lambda_mean*dose.range))*dose.range)
lines(dose.range.orig, y, lwd = 2)
# add upper and lower 90% credible interval for curve
pci.ul <- pci.ll <- numeric(length(pred.doses))
for(ii in 1:length(pred.doses)) {
# quantiles <- quantile(unlist(jags.samples$y.pred[ii, , ]), probs = c(0.025, .975))
quantiles <- quantile(unlist(jags.samples$mu.tilde[ii, , ]), probs = c(0.05, .95))
pci.ll[ii] <- quantiles[1]
pci.ul[ii] <- quantiles[2]
}
lines(dose.range.orig, pci.ll, col = 'grey', lwd = 2)
lines(dose.range.orig, pci.ul, col = 'grey', lwd = 2)